Brain Tumor Segmentation Algorithm Based on Feature Fusion and Attention Mechanism

Abstract

The accurate segmentation of brain tumors in Magnetic Resonance Imaging(MRI)is of great significance for surgical planning and radiotherapy. U-Net, as the most widely used network in the field of brain tumor segmentation, has excellent performance; however, there are problems such as significant semantic differences in skip connections and insufficient utilization of cross channel information in MRI images. To accurately segment various regions of brain tumors, an improved U-Net model, Feature Fusion(FF)and attention mechanism FFCA-U-Net is proposed. The FF module is designed in skip connections to replace the direct splicing operation in U-Net, to effectively fuse feature information at different levels and scales, reduce semantic differences, adjust receptive fields, and enhance the network's learning ability for tumor features. An improved three-dimensional Coordinate Attention(CA)mechanism is introduced into the encoder to capture cross channel information along the three directions of the MRI image, enhancing the network's perception of brain tumor boundary information, thereby obtaining more accurate positions of tumor subregions. In addition, to readily obtain the relative position of tumors and reduce network learning redundancy, mask images are added alongside MRI images as network inputs. In the experimental results on the MSD dataset, the Dice coefficients of FFCA-U-Net in the enhanced tumor area, non-enhanced tumor area, and edema area are 0.803 4, 0.628 6, and 0.799 3, respectively, with an average Dice of 0.743 8, which is superior to those of other advanced networks such as TransBTS and UNETR.